脑子里那个声音

不停嘴地念「the the the」，一串字母就记不住了

先做一件事。看这七个字母，遮住，再按顺序默写：B C D G P T V。多数人会串位、会漏。换成 F K L Q R Y S，同样七个，明显好办。差别只在读音：前一串念出来都带同一个"衣"的尾音，后一串各不相同。Conrad 和 Hull 1964 年报告的这个现象叫语音相似性效应——关键是你是用眼睛看的、全程没出声，可记不住的偏偏是读音黏在一起的那串：你存它们的时候把它们转成了声音。这效应不挑语言，你可以自己试：「四 十 是 事 时 使」和「猫 车 树 灯 门 伞」，哪串更容易串位。

第二件事更直接。让人一边不停地小声念「the the the the」，一边记那串字母——跨度掉下来，而且语音相似性效应跟着消失了：两串变得一样难记。这个手法叫发音抑制（articulatory suppression）。它占住的不是你的注意力，是"在心里念一遍"这个动作本身。

拼起来就是 Baddeley 那套语音回路（phonological loop）：一小块会自己褪掉的声音存储（大约两秒的量），加一个靠"心里念一遍"给它续命的排练机制。脑子里那个声音到这里第一次有了一份具体的工作：它不是思维的伴奏，它在刷新一块快要褪掉的东西。

如果这是真的，那"读起来更费时间的语言，一口气能刷住的就更少"——这是个能查的预言。Ellis 和 Hennelly 1980 年在威尔士语—英语双语者身上查了它：威尔士语的数词念出来平均比英语的长（"四"是 pedwar，两个音节，英语 four 只有一个），同一批人用威尔士语记数字串的跨度就明显短于用英语。他们的本意恰恰是反对把数字跨度当智力指标：差的不是脑子，是每秒能刷几个。

但这块地没夯实。Stigler、Lee 和 Stevenson 1986 年发现汉语儿童的数字跨度比美国、日本同龄人平均多两位以上，而汉语的一到十全是单音节，看着正好接上；可 Mattys、Baddeley 和 Trenkic 2018 年实测双语者的发音速度，发现它撑不起这个差距——"跨度随发音速度走"的相关在英语母语者身上有，在汉语母语者身上弱得多。所以这里也别读成哪门语言"更适合算数"：数字跨度只是一种测法，双语者内部对照（威尔士语那份）比跨国比较硬得多。

EN Cover and recall seven letters in order: B C D G P T V is much harder than F K L Q R Y S. The first set rhymes; the second does not. This is the phonological similarity effect (Conrad & Hull, 1964), and the point is that the letters were seen, not heard — so they were stored as sound. Now have someone mutter "the the the" while doing the task: span drops and the similarity effect disappears. That technique, articulatory suppression, occupies the act of saying things silently to yourself. Together they define Baddeley's phonological loop: a sound store that fades in roughly two seconds, plus a rehearsal mechanism that refreshes it. Inner speech thereby gets a job description — it is the refresh. Prediction: a language whose words take longer to say should let you refresh less. Ellis and Hennelly (1980) found exactly that in Welsh–English bilinguals, whose digit span was shorter in Welsh, where number words take longer to articulate — their point being that Welsh digit-span norms reflect articulation, not intelligence. Stigler, Lee and Stevenson (1986) found Chinese children recalled at least two more digits than American or Japanese peers, and Chinese number words are monosyllabic. But Mattys, Baddeley and Trenkic (2018) measured articulation speed directly and found it inadequate to explain the Mandarin–English gap: the span–speed correlation held for English natives and was much weaker for Mandarin natives. Within-bilingual comparisons are the stronger evidence; the mechanism is still disputed. No language here is "better for arithmetic" — digit span is one measure among others.

它是压缩的、会跟你吵架，而且未必是声音

Hurlburt 用了个笨办法：给人戴一个随机响的耳机，响的那一下把"响之前最后一瞬意识里有什么"记下来，第二天再逐条面谈追问。这叫随机时刻抽样（descriptive experience sampling）。Heavey 和 Hurlburt 2008 年的结果：30 个人、每人 10 个随机时刻，内语（inner speech，脑子里那个不出声的说话）出现在大约四分之一的时刻。

另外几类的占比也各是四分之一上下：脑子里的画面、情绪感受、感官觉知，还有一类叫"无符号思考"——你明确在想一件确定的事，那一瞬却既没有词也没有画面。人和人差得很开：有人几乎每次响铃都在心里说话，有人一次都没有。这份证据样本小、靠自述、访谈技术本身还是争论对象，但它足够掀翻一件事——"我一直在心里说话"这个印象不可靠，你记得的只是你注意到的时刻。

形状也不像朗读。Fernyhough 用问卷把内语拆成几个维度：压缩还是展开、独白还是对话、有没有别人的声音、是不是在评价你。压缩最容易自己验：心里那句常常只剩一个谓语，主语和背景全省掉——"完了"、"再往左一点"，硬要展开成完整句子反而费劲。对话那一维是反刍的地基：你在心里替对方把话说了，这段就能自己接下去。（孩子解题时的出声自语也是这样：任务越难自语越多，维果茨基认为内语就是它被"搬进去"的结果。）

最有意思的是手语者。如果那个声音真是"声音"，聋人手语者就该没有内语；实际上他们报告在心里"打手"。Wilson 和 Emmorey 1997 年把语音相似性效应换模态去测：给聋人手语者看一串手语，手形相近的串比手形不同的串更难按顺序记住；1998 年又测出手语长度效应——动作路径长的手语串更难记，对应口语的词长效应。同一个回路，排练的是手不是嘴。

所以"脑子里那个声音"只说对了一半。更准的说法是：你那门语言的形式正在被排练——是嘴还是手，取决于你会哪门语言。

EN Hurlburt's descriptive experience sampling gives people a beeper that fires at random moments; at the beep they note what was in awareness just before it, then get interviewed about each moment. Heavey and Hurlburt (2008) sampled 10 random moments from each of 30 people: inner speech showed up in roughly a quarter of them, as did inner seeing, feelings, sensory awareness, and "unsymbolized thinking" — clearly thinking something definite with neither words nor images present. Individual variation was wide. The evidence is weak in the usual ways (small sample, self-report, contested interview method) but strong enough to unseat the impression that you are talking to yourself all the time. Fernyhough's questionnaire work splits inner speech into dimensions: condensed or expanded, monologic or dialogic, other people's voices present or not, evaluative or not. Condensation is easy to check yourself: the inner sentence is often a bare predicate with the subject dropped. Children give the developmental evidence: private speech rises with task difficulty and fades on easy tasks, and children who use more of it are more attentive and more successful on hard tasks — a correlation, not a recipe. And inner speech is not tied to sound. Wilson and Emmorey (1997) found that deaf signers recall lists of handshape-similar signs worse than dissimilar ones, and (1998) that signs with longer movement paths are recalled worse — the similarity and length effects, transposed to the hands. Same loop; the rehearsal is manual.

有人报告根本没有那个声音，而他们并没有不能思考

Nedergaard 和 Lupyan 2024 年做了第一批系统对照：用自评量表把人分成内语体验多的 47 人和少的 46 人，上四个任务。差别出现在两个任务上：记一串读音相近的词（就是上面那个语音相似性的活），内语少的那组更差；看两张图判断它们的名字押不押韵（一只猫和一顶帽子，英语里 cat 和 hat 押韵），内语少的那组更吃力。

另外两个任务没差别：任务切换（按提示在两套规则间来回切，一般认为要靠心里念着当前规则）和知觉判断上的范畴效应。作者给"没有内语体验"起了个名字：anendophasia（无内语）。

证据强度：单个研究、93 人、分组靠自评。Lind 2025 年在同一刊物上评论说，结果本身是新贡献，但"有人完全没有内语"证据不够——量表问的是"你体验到多少"，不等于"你用到多少"。目前站得住的只有一句：自述的内语多少，与某些言语任务的表现相关。

更硬的一刀来自失语症。Varley、Klessinger、Romanowski 和 Siegal 2005 年报了三位左半球语言区大面积损伤的男性：语法严重受损、连读写数词都困难，可基本的计算程序完整——四则运算的结构、哪一步该先算，都在。同一组更早的工作还发现，语法严重受损不妨碍通过心智理论题（判断一个抱着错误想法的人接下来会怎么做）。语法这套机器坏了，算术和推理的骨架没跟着坏。

于是这场争论今天大致在这里：强版本（没有语言就没有命题思维）基本没人持有了，活下来的是弱版本的两支。一支说语言是增强器，它让某几件事变容易——把一串东西按顺序存住、给自己下指令、把注意力钉在一个范畴上；拿掉它，掉的正是这几样，不是思维本身。另一支更强，说某些概念操作离了语言标签根本做不出来。anendophasia 那份研究的结果恰好卡在两支中间：言语任务掉了，任务切换和范畴效应没掉。

最后把话放正：报告自己没有内心独白，既不是缺陷也不是天赋——这一行连怎么测它都还在讨论。

EN Nedergaard and Lupyan (2024) ran the first systematic comparison: a self-report questionnaire split participants into high (47) and low (46) inner-speech groups, then four tasks. Two showed a difference — verbal working memory for lists of similar-sounding words, and judging whether two pictured objects' names rhyme (a cat and a hat) — and two did not: task switching, previously tied to covert verbal cueing, and categorical effects on perceptual judgements. They named the absence of inner-speech experience anendophasia. Evidence strength: one study, 93 people, grouped by self-report. Lind (2025) replied in the same journal that the results are a genuine contribution but that no compelling evidence shows anyone wholly lacks inner speech: the scale asks how much you experience, not how much you use. The harder blow to the strong view comes from aphasia. Varley, Klessinger, Romanowski and Siegal (2005) described three men with large left perisylvian lesions: grammar severely impaired, number words hard to read or write, and yet all basic computational procedures intact. Earlier work from the same group found severe grammatical impairment did not preclude passing false-belief tasks. Grammar broke; the skeleton of arithmetic and reasoning did not. So the strong claim that thought is inner speech has few defenders now. What survives are two weaker positions: language as an enhancer of specific operations (holding an ordered list, cueing yourself, pinning attention to a category), and the stronger claim that some conceptual operations cannot run without linguistic labels at all. The anendophasia results sit between them. Reporting no inner monologue is neither a deficit nor a gift.